NBER WORKING PAPER SERIES

ON ESTIMATING THE EXPECTED RETURN ON

THE MARKET: AN EXPLORATORY INVESTIGATION

Robert C. Merton

Working Paper No. 444

NATIONAL BUREAU OF ECONOMIC RESEARCH

1050 Massachusetts Avenue
Cambridge MA 02138

February 1980

The research reported here is part of the NBER's research

project on Debt and Equity. Any opinions expressed are
those of the author and not those of the National Bureau
of Economic Research.
 NBER Working Paper 444
February, 1980

On Estimating the Expected Return on the Market:

An Exploratory Investigation

ABSTRACT

The expected return on the market is a number frequently required for

the solution of many investment and corporate finance problems. However,
by comparison with other financial variables, there has been relatively little
academic research on estimating this expected return. The current practice
for estimating the expected return on the market is to take the historical
average of realized excess returns on the market and add, :It to'the current
observed inte're~t: rate. While th;i:smodel does explicitly relfect the depen-
dence of the market experience return ,on the interest rate, it does not
take into account the effect of changes in the level of risk associated
with the market.
Three models of equilibrium expected market returns which do reflect
this dependence are analyzed in this paper. Estimation procedures are
derived which incorporate the prior restriction that equilibrium expected
excess returns on the market must be positive. The parameters of the models
are estimated using realized return data for the period 1926-1978.
The principal conclusions from this exploratory investigation are:
~irst, in estimating models of the expected return on the market, the non-
negativity restriction of the expected excess return should be explicitly
included as part of the specification. Second, estimators which use
realized return time series should be adjusted for heteroscadasticity.

Robert C. Merton
Department of Economics
Massachusetts Institute of Technology
Bldg. E52, Room 243
Cambridge, Massachusetts 02139

(617) 253-6617
 ON ESTIMATING THE EXPECTED RETURN ON THE MARKET:
An Exploratory Inv~stigation
Robert C. Merton*
Massachusetts Institute of Technology

I. Introduction

Modern finance theory has provided many insights into how security

prices are formed and has provided a quantitative description for the

risk structure of equilibrium expected returns. In the most basic form

of the Capital Asset Pricing Model,ll this equilibrium structure is given

by the Security Market Line relationship: Namely,

ai - r • ai (a - r) (1.1)

where a and a denotes the expected rate of return on security i and

i
the market portfolio, respectively; r is the riskless interest rate; and

ai is the ratio of the covariance of the return on security i with the

return on the market divided by variance of the return on the market.

,This same basic model tells us t~at all efficient or optimal portfolios

can be represented by a simple combination of the market portfolio with

e
the riskless asset. Hence, if ae and cr are the expected rate of

return and standard deviation of return on an efficient portfolio, then

ae = w(a - r) + rand cre = we where w· is the fraction allocated to

the market and cr is the standard deviation of the return on the

market. From these conditions, we have that the equilibrium tradeoff

between risk and return for efficient portfolios is given by

(1.2)
 -2-

(1.2) is called the Capital Market Line and (a - r)/a, the slope of

that lin~is called the Price of Risk.

From (1.1) and (1.2), one can determine the optimal portfolio

allocation for an investor and the proper discount rate to employ for the

evaluation of securities. Moreover, these equations provide the critical

"cost of capital" or "hurdle rates" necessary for corporate capital

budgeting decisions. Of course, (1.1) and (1.2) apply only for the most

basic version of the CAPM, and indeed, empirical tests of the Security

Market Line have generally found that while there is a positive relation-

ship between beta and average excess return, there are significant

deviations from the predicted relationship.!/ However, these deviations

appear principallyin very "high" and very "low" beta securities.

3
MOreover, there is some question about the validity of these tests. /

The more sophisticated intertemporal and arbitrage-model versions of the

. 4/
CAP~ show that equilibrium expected returns on securities may depend

upon other types of risk in addition to "systematic" or "market" risk,

'arid hence, they provide a theoretical foundation for (1.1) and (1.2) not

to obtain. However, in all of these models, the market risk of

a security will affect its equilibrium expected return, and indeed, for
. 5/
most common stocks, market risk will be the dominant factor.-

Thus, at least for common stocks and broad-based equity portfolios, the

basic model as described by (1.1) and (1.2) should provide a reasonable

"first approxi.Jnation" theory for equilibrium expected returns.

Of course, all one needs to know to apply these formulas in solving

portfolio and corporate financial problems are the parameter values.

 -3-

And as might be expected, considerable effort has been applied to

estimating them. However, this effort has not been uniform with respect

to the different parameters, and as will be shown, this nonuniformity is

not without good reason.

For the most part, r is an observable, and so that parameter is

gotten for free. Among the other parameters, beta is the one most widely

estimated. In dozens of academic research papers, betas have been

estimated for individual stocks; portfolios of stocks; bonds and other

fixed income securities; other investments such as real estate; and

even human capital..§.! For practitioners, there are beta "books" and

beta services. While for the most part these betas are estimated from

time series of past returns, various accounting data have also been used.

In their pioneering work on the pricing of options and corporate

liabilities, Black and Scholes (1973) deduced an option pricing formula

whose only nonobservable input is the variance rate on the underlying stock.

As a result, there has been a surge in research effort to estimate the variance

. rates for returns on both individual stocks and the market. Although this

research activity is still in its early stages of development, variance rate

estimates are already available from a number of sources.

In contrast, there has been little academic research on

estimating the expected return on either individual stocks or the market.

Ibbotson and Sinquefield (1976; 1979) have carefully cataloged the

historical average returns on stocks and bonds from 1926 to 1978.

However, they provide no model as to how expected returns change through

time. There is no literature analogous to the term structure of

 -4-

interest rates for the expected return on stocks, although there is

research going on in this direction as, fOr example, in Cox, Ingersoll,

and Ross (forthcoming).

One possible explanation for this paucity of research on expected

returns is that for many applications within

finance, only relative pricing relationships are used, and therefore,

estimates of the expected returns are not required. Some important

examples of such applications are option and corporate liabilities pricing

and the testing for superior performance of actively-managed portfolios.

However, for many if not most applications, an estimate of the expected

return on the market is essential. For example, to implement even the

most passive strategy of portfolio allocation, an investor must know

the expected return on the market and its standard deviation in order to

choose an optimal mix between the market portfolio and the riskless asset.

Indeed,· even if one has superior security analysis skills so that the

optimal portfolio is no longer a simple mix of the market and the riskless

asset, TreynJr and Black (1973) have shown that the optimal strategy will

still involve mixing the market portfolio with an active portfolio, and

the optimal mix between the two will depend upon the expected return and

standard deviation of the market. For a corporate finance example, the

application of the model in determining a "fair" rate of return for

investors in regulated industries requires not only the beta but also an

estimate of the expected return on the market. As these examples

illustrate, it is not for want of applications that expected return

estimation has not been pursued.

A more likely explanation is simply that estimating expected returns

from time series of realized stock return data is very difficult. As is

shown in Appendix A, the ·estimates of variances or covariances from the

 -5-

available time series will be much more accurate than the. corresponding

expected return estimates. Indeed, even if the expected return on the

market were known to be a constant for all time, it would take a very long

history of returns to obtain an accurate estimate. And, of course, if

this expected return is believed to be changing through time, then

estimating these changes is still more difficult. Further, by the

Efficient Market Hypothesis, the unanticipated part of the market return

(i.e., the difference between the realized and expected return) should not

be forecastab1e by any predetermined variables. Hence, unless a significant

portion of the variance of the market returns is caused by changes in the

expected return on the market, it will be difficult to use the time

series of realized market returns to distinguish among different models

for expected return.

In light of these difficulties, one might say that to attempt to

estimate the expected return on the market is to embark on a fool's

errand. Perhaps, but on this errand, I present three models of expected

. return and derive methods for estimating them. I also report the results

of applying these methods to market data from 1926 fo 1978.

The paper is exploratory by design, and the empirical estimates

presented should be viewed with that in mind. Its principal purpose is

to motivate further research in this area by pointing out the many estima-

tion problems and suggesting directions for possibly solVing them. The

reasons for taking this approach are many: First, an important input for

estimating the expected return on the market is the variance rate on the

market. While there is much research underway in developing variance

 -6-

estimation models, their development has not yet reached the point where

there is a "standard" model with well-understood error properties.

Because this is not a paper on variance estimation, the model used to

estimate variance rates here is a very simple one. Almost certainly,

these variance estimates contain substantial measurement errors, and these

alone are enough to warrant labeling the derived model estimates for

expected return as "preliminary." A second reason is that the expected

return model specifications are themselves very simple, and undoubtably

could be improved upon. Third, only time series data of market returns

were used in the estimations, and as is indicated in the analysis, other

sources of data could be used to improve the estimates. As a reflection

of the preliminary nature of this investigation, no significance tests are

provided and no attempt is made to measure the relative forecasting

performance of the three models.

 -7-

II. The Models of Expected Return

The appropriate model for the expected return on the market will

depend upon the information available. For example, in the absence of

any other information, one might simply use the historical sample average

of realized returns on the market. Of course, we do have other informa-

tion. For example, we can observe the riskless interest rate. Noting

that this rate has varied between essentially zero and its current doub1e-

digit level during the last fifty years, we can reject the simple sample

average model for two reasons: First, it can be proved as a rather

general proposition that a necessary condition for equilibrium is that

the expected return on the market must be greater than the riskless rate

(i . e., a » r. 11 Hence, if the current interest rate exceeds the long

historical average return on stocks (as it currently does), then the

sample average is a biased-low estimate. Thus, one would expect the

expected return on the market to depend upon the interest rate. Second,

the historical average is in nominal terms, and no sensible model would suggest

. that the equilibrium nominal expected return on the market is independent

of the rate of inflation which is also observable. Both these criticisms

are handled by a second-level model which assumes that the expected excess

return on the market, a - r, is constant. Using this model, the current

expected return on the market is estimated by taking the historical average excess

return on the market and adding to it the current observed interest rate.

Indeed, a model of this type represents essentially the state-of-the-art

with respect to estimating the expected return on. the market •.§.1

This model explicitly recognizes the dependence of the market

 -8-

expected return on the interest rate and in so doing, it implicitly takes

/-

iuto account the level of inflation. However, it does not take into account

another important determinant of market expected return: Namely, the level

of risk associated with the market. At the extreme where the market is risk-

less, -then by arbitrage, a. = r, and the risk premium on the market will be

zero. If the market is not riskless, then the market mu~t have a positive risk

premium.
9/
While it need not always be the case,- a generally-reasonable

assumption is that to induce risk-averse investors to bear more

risk, the expected return must be higher. Given tha~ in the aggregate,

the market must be held, this assumption implies that, ceteris paribus, the

equilibrium expected return on the market is an increasing function of the

risk of the market. Of course, if changes in preferences or in the

distribution of wealth are such that aggregate risk aversion declines

between one period and another, then higher market risk in the one

period need net imply a correspondingly higher risk premium. However, if

aggregate risk aversion changes slowly through time by comparison with

- the changes in market risk, then, at least locally in time, one would -

expect higher levels of risk to induce a higher market risk premium.

If, as shall be assumed, the variance of the market return is a

sufficient statistic for its risk, then a reasonably general specification

of the equilibrium expected excess return can be written as

2
a. - r .. Yg(O' ) (11.1)

2
where g is a function of 0'2 only, with g(O)" 0 and dg/dO' > O.

In the analysis to follow, we shall assume that the function g is

 -9-

2
known and that 0 can be observed. It is also assumed that there is a
2
set of state variables S in addition to the current 0 that can be

observed. The specific identity of these state variables will depend upon

the data set available. However, Y is not one of these state variables.

Hence, conditional on this information set, the expected excess return on

the market is given by

(11.2)

where E[ Is,a2 ] is the conditional expectation operator, conditional

on knowing S and o2 • Since Y is not observable, for (11.2) to

have meaningful content, the further condition is imposed that

(11.3)

That is, given the state variables S, the conditional expectation of Y

does not depend upon the current o2 . This condition, of course, does
2
not imply that Y. is independent of o . Thus, from (11.3), we can

rewrite (11.2) as

(11.4)

Since it has already been assumed that variance is a sufficient statistic

for risk, with little loss in generality, it is further assumed that

market returns can be described by a diffusion-type stochastic process

101
in the context of a continuous-time dynamic mode1. Specifically, the

instantaneous rate of return on the market (including dividends), dM/M,

can be represented by the Ita-type stochastic differential equation

dM(t) • + odZ(t) (11.5)

M(t)
adt
,.
•
 -10-

where dZ(t) is a standard Wiener process and (11.5) is to be inter-

preted as a conditional equation at time t, conditional on the ins tan-

taneous expected return on the market at time t, a(t) =a and on the

instantaneous standard deviation of that return at time t, a(t) = a.

Und er " d "~t10ns,--
certa~n "
con 11/"~t can b e Slown
1 in the context of

an intertemporal equilibrium mod~l that toe equilibrium instantaneous

expected excess return on the market can be reasonably approximated by

(11.6)

where Yl is the reciprocal of the weighted sum of the reciprocal of each

investor's relative risk aversion and the weights are related to the distri-

bution of wealth among investors. To add further interpretation for

Y ,
l
in the frequently-assumed case of a representative investor with a constant

relative risk aversion utility function, Y would be an exact constant

l
and equal to this representative investor's relative risk aversion. The specifi-

cation for expected excess return given by (11.6) which will be referred

to as "Model 111" is indicative of models where it is assumed that aggre-

gate risk preferences remain relatively stable for appreciable periods of

time.

"Model In" makes the alternative assumption that the slope of the

Capital Market Line or the Market Price of Risk remains relatively stable

for appreciable periods of time. Its specification is given by

(11.7)
 -11-

where Y2 is the Market Price of Risk. Like "Model 111," it allows for

cl,anges in the expected excess return as the risk level for the market changes.

"Model #3" is the state-of-the-art model which assumes that the

expected excess return on the market remains relatively stable for

appreciable periods of time even though the risk level of the market is

changing. Its specification is given by

(11.8)

Of course. if the variance rate on the market were to be essentially

constant through time, then all three models would reduce to the state-of-

the-art model with a constant expected excess return. However, from the

work of Rosenberg (1972) and Black (1976) as well as many others, the

hypothesis that the variance rate on the market remains constant over any

appreciable period of time can be rejected at almost any confidence level.

MOreover, given that the variance rate is changing, the three models are

mutually exclu~ive in the sense that if one of the models satisfies

condition (11.3), then the other two models cannot. To this, note that
_ j-i
Y - Yi[a(t}] for i,j = 1,2,3. Therefore, if Y satisfies (11.3),
j i
then E[Yjls] = E[Yils]E{[a(t)]j-ils}. E[Yjls,a2(t)] = E[Yils] [a(t)]j-i.
Therefore, for i + i, Y
j
can only satisfy (11.3) if

E{[aCt)]j-iI S } = [a(t)]i-i for all possible values of aCt), and this

is not possible unless the {a(t)}· are constant over time.

While we have assumed that ~(t) is observable, in reality, it is

not, and therefore; like aCt), it must be estimated. Hence, these

models as special cases of (11.1) will be of empirical significance only

 -12-

if for the available data set, the variance rate can be estimated more

accurately than the expected return. If the principal component of such

a data set is the time series of realized market returns, then it is

shown as a theoretical proposition in Appendix A that, indeed, the

variance rate can be more accurately estimated when the market return

dynamics are given by (11.5). As an empirical proposition, the studies

of both Rosenberg (1972) and Black (1976) show that a nontrivial portion

of the change in the variance can be forecasted by using even relatively

simple models. Further, along the lines of Latane' and Rendleman (1976)

and Schmalensee and Trippi (1978), it is possible to use observed option

prices on stocks to deduce "ex-ante" market estimates for variance rates

by "inverting" the Black-Scholes option pricing formula. Hence, models

of the type which satisfy (11.4) hold out the promise of better estimates

for the expected return on the market than can be obtained by direct

estimation from the realized market return series.

While (11.5) describes the dynamics of realized market returns,

we have yet to specify how aCt) and Y., j = 1,2,3 change through
J
time. Although aCt) changes through time, it is assumed to be a
slowly-varying function of time relative to the time scale of market

price changes, and,'therefore, over short intervals of time,' the variation- in

realized market returns will be very much larger than the variation in

the variance rate. That is, it is assumed that

for satisfactorily small 0, there exists a finite time interval h such

2 2
that the probfla (s) - a (t) I > o!EI£(t,t + h)} will be essentially zero
where ,,2(t) " [1.' ,,2(s)dslh. In essence, we assume that the variance
 -13-

rate can be treated as constant over finite time intervals of length h

and that h» dt. In a similar fashion, it is also assumed that the

riskless interest rate can be treated as constant over this same finite

time interval h.

Under the hypothesis that Model OJ is the correct specification, we

assume that Y is a slowly-varying function of time relative to the

j
time scale of changes in the variance rate. That is, there exists a

finite time interval T, T» h, such that Y can be treated as

j
essentially constant over intervals of that length. Again, because

Y = yi[a(t)]j-i, i, j
j
= 1,2,3 if one of the models satisfies this

assumption, then the other two cannot.

It follows immediately from these hypothesized conditions and the

model specifications that the expected rate of return on the market,

aCt), can be treated as essentially constant over time intervals of

length h. Therefore, over short intervals of time, the variation in the

expected return on market will be similar in magnitude to the variations

2
. in cr (t) and r(t) and very much smaller than the variation in

realized market returns.

Let ~(t):: Met + h}!M(t) denote the return per dollar on the

market portfolio between time t and t + h. Under the hypothesized

conditions for the dynamics of a(t) and Y , we have from (II.5) that
j
conditional on knowing M(t), a(t), and ret), ~(t) will be lognormally

distributed. [t+h ]
Let R(t):: exp ~ r(s)ds denote the return per dollar on the

riskless asset between t and t + h and define X(t) :: In[~(t)!R(t)].

 -14-

Under the hypothesis that Model OJ is the correct specification~ we can

e:..:press X(t) a ;g./

'"

(II. 9)

t+h
where Z(t;h) - J( dZ is a normally distributed random variable with

mean equal to zero and a standard deviation equal to~. Moreover, for

all t and t' such that It' - tl ~ h, Z(t;h) and Z(t';h) will be

independent.

In preparation for the model estimation, we proceed as follows:

Let T denote the total length of time over which we have data. The

first step is to partition the data into neT)(= T!T) nonoverlapping

time periods of length T. By hypothesis, will be constant within

each of these neT) time periods. The second step is to partition each

of these neT) time periods into N(= T!h) nonoverlapping subperiods of

length h. By hypothesis, the variance and interest rates will be

-constant within each of these N subperiods.

Since by hypothesis none of the variables relevant to the estimation

changes during any of the nonoverlapping subperiods of length h, there

is nothing to be gained by further subdivisions. Hence, the interval

between observations will equal h, and by appropriate choice of time

units, h can be set equal to one. Therefore, all time-dimensioned

variables are expressed in units of the chosen observation interval.

Because within each of the neT) time periods, the subperiods are

of identical length and nonoverlapping, it should cause no confusion to

 -15-

th '
redefine the symbol "t" to mean "the ~ subperiod of length' h" within

a particular time period of length h. So redefined, t will take on

integer values running from t = 1, ••• ,N. There is no need to further

distinguish "t" as to the time period in which it takes place because

(a) the posited stochastic processes are time homogeneous; (b) the

length of the subperiods are the same for all n(T) time periods; and

(c) the neT) time periods are nonoverlapping. By the choice for time
th
units, "t" will also denote the ~ observation within a particular

time period.

With t redefined and h = 1, (II.9) can be rewritten for a

particular time period as

t = 1, ••• ,N (IL10)

where g(t) is a standard normal random variable. Because the subperiods

are nonoverlapping, g(t) and g(t') will be independent for all t and

t' such that t ~ t'. For the N observations within this time period,

Y is, by hypothesis, a constant.

j
With this, the descriptions of the models are complete, and we now

turn to the development of the estimation procedures.

 -16-

III. The Estimation Procedures

Given a time series of estimates for aCt), the natural estimation

procedure suggested by (11.10) is least-squares regression. (11.10) is put

in standard form, by making the change in variables X'{t) = x{t)!a(t)

+ a{t)!2 and rewriting (11.10) for Model OJ as

X'(t) = Yj [a{t)]2-j + E(t)

.
(II1.l)

Given the N observations within the time period over which Yj is

constant, we have that the least-squares estimator for Y , Y'" ,

j j
j = 1,2,3, can be written as

"'.
Model II: Y
1
= (II1.2.l)

Model 12: '" = {Nt[X{t)!a(t)]+.S

Y2 (II1.2.2)

(III.2.3)

From (111.1), all the conditions for least-squares are satisfied, and

therefore, '"
Y appears to be the best linear unbiased estimator ofY .
j j
Since realized rates of return on the market can be negative, it is certainly

possible that for a particular time period, Y could ·be negative. In such a
j
case, is that value for Y an unbiased estimate of Y ? From prior knowledge,
j j
aCt) ~ ret) must be positive. Therefore, each of the Y must be
j
positive, and the answer to the question is "no." Thus, (111.1) is not a
 -17-

complete description of Model #j's specification. A complete description

must include the condition Y > O.

j
While there are a variety of ways to incorporate this restriction,

it is done here by assuming a prior distribution for Y and applying

j
Bayes' Theorem to deduce a posterior distribution based upon the observed

data. The specific prior chosen is the uniform distribution so that the prior

density for Y
j
is given by f(Y )
j
= lIb where 0 ~ Y
j
~ b.

Conditional upon knowing Y

j
and cr(t), t = 1, ••• ,N, we have

from (111.1) that the X'(t), t = 1, ••• ,N are independent and joint

normally distributed. Using the uniform prior assumption for Y., it

J
is shown in Appendix B that the posterior density function for Y ,
j
F[YjIX'(t),cr(t),t = 1, ••• ,N], will satisfy j = 1,2,3,

(111.3)

where ~() is the cumulative standard normal density function;

., (111.4.1)

N
rl _ L [cr(t) ]4-2j ., (III. 4. 2)
j 1

Pj == OJ (b - Aj ) and nj == -AjOj •
By inspection of (111.3) and (111.4), the way in which the data enter

the posterior distribution can be summarized by two statistics: A and

j
2
OJ. Moreover, by comparing (III.4.1) with (111.2), we have that

(111.5)
 -18-

To reflect these observations t the posterior distribution is written as

" 2
F[Y IY tQ ;b]. Further inspection of (III. 3) will show that F is a
j j j
truncated normal distribution on the interval [Otb] with characteristic

parameters Y and
j
As Figure 111.1 il1ustrates t the posterior density function will be a

monotonically decreasing function of Y if Y ~ 0 and a monotonically

j j
'" "
increasing function if Y ~ b.lf 0 < Y < b t then F monotonically
j j
" "
increases for 0 ~ Yj < Yj; reaches a maximum at Y = Y ; and monotoni-
j j
cally decreases for " < Y. < b.
Y. It follows immediately that the
J J -
maximum likelihood estimate of Y based upon the posterior distribution t
j
YR. will satisfy
r
yR. = 0 for " <0
Y
j j

"
=Y
j
for o ~ Y" j <b (111.6)

= b for " >b

Y
j

However t for the purposes of this analysist the maximum likelihood

estimator is not the proper choice. The objective is to provide an

estimate of Y for the prediction of the expected excess return on the

j
2
market t conditional on knowing the current variance rate t cr (t).

Conditional upon Model Dj being the correct specification t we have

from (11.3) and (II.10)t that

FIGURE: lII.1 POSTERIOR. l)fNS,ry FUNCTION

. .. A

. . . . ..... ..)J 4: 0
..._~
...... . .=0
.

y..)
o b

o b

)J >b
._ .Y:=b

y.
J
o b
 -19-

E[a(t) - ret) la2 (t) ,S] = E[X(t) + .5a2 (t) Icr2 (t) ,S]

= [cr(t)]3-jEIYjlcr2(t),S] (111.7)

• [C1(t)]3- j E[Y IS]

j

where in this context, S denotes the set of data available to estimate

the distribution for Y.. From (111.7), it therefore follows that the
J
correct estimator to use for estimating the expected excess return is the

expected value of Y. computed from the posterior distribution.

J

_ "2
As is derived in Appendix B, Y
j
= E[Yj Iyj,nj;b], j = 1,2,3, is

given by

(111.8)

where Yj == bIz is the expected value of Y

j
based upon the prior

distribution.
." i
From (111.6) and (111.8), the relationship between Y , Y , and Y
j j j
for a finite number of observations can be summarized as follows:
 -20-

"
Y > yR. > Y for "
Yj ~ 0
j j j

Y > yR. =~ for o ~ Y" j < b/2

j j j

R. " "
Y = Y = Y
j j j
for Yj = b/2 = Y
j
(111.9)

Y < yR. =
j j
ij for b/2 < Y
A

j
~b

Yj < Y~ < ij for

"
Y >b
j

If the model is correctly specified so that in the limit as the number of

observations N becomes large, Y converges to a point in the interval

j
R. "
(O,b], then Y. = Y., and from (111.8), converges to Y•• Hence, both
J J J
are consistent estimators.

established the model estimator properties, we now turn

\
. to the estimation of the models.
 -21-

IV. Model Estimates: 1926 to 1978

In this section, market return and interest rate data from 1926 to

1978 are used to estimate each of the three models presented in Section

II. The model estimators are the ones derived in Section III. The

monthly returns (including dividends) on the New York Stock Exchange

Index are used for the market return series. This index is a value-weighted

portfolio of all stocks on the New York Stock Exchange. The U.S. Treasury

Bill Index presented in Ibbotson and Sinquefield (1979) is used for the

~iskless interest rate series. The monthly interest rate from this

index is not the yield, but the one-month holding period returns on the

shortest maturity bill with at least a thirty-day maturity.

The interval h over which it is assumed that the variance rate on

the market can be treated as constant was chosen to be one month. The

riskless interest rate is also assumed to be constant during this interval,

and one month is, therefore, the observation interval. The choice of

a one-month interval w~s\certainly influenced by the availability of ~ata.

However, a one-month interval is not an unreasonable choice. At least

in periods in which daily return data are available, this interval is

long enough to permit reasonable estimates of the variance rate along

the lines discussed in Appendix A, and it is short enough so that the

variation in the variance rate over the observation interval is

substantially smaller than the variation in realized returns.

Other than satisfying the condition that T be significantly

larger than h, I have no a priori reasons for choosing a specific

value for the length of the time period over which Y is assumed to be
j
 ..22..

constant. Perhaps other data besides market returns would be helpful.

For example, if the data on large samples of individual investors'

holdings of various types of assets such as those used in the Blume and

Friend study (1975) were available for different points in time, it

might be possible to use these data to estimate the changes in aggregate

relative risk aversion over time. However, given the explorato.ry nature

of this investigation, the route taken here is simply to estimate the models

assuming different values for T ranging from one year to fifty-two years

and to examine the effect of these different choices on the model estimates.

A third choice to be made is the value to assign to b in the

uniform prior distribution for Y • Unlike the lowerbound nonnegativity

j
restriction on Y., there are no strong theoretical foundations for an
J
upperbound on relative risk aversion, and therefore, for an upperbound on

equilibrium expected return3. For b to be part of a valid prior, the

market return data used to form the posterior cannot be used to form an

empirical foundation for/,the upperbound restriction. Again, estimates of

aggregate risk aversion from the investor data used in the previously-

cited Blume and Friend study might provide some basis for setting b.

However, in the absence of such other information, a reasonable choice

is a diffuse prior on the nonnegative real line with b = 00. Taking the

limit as b goes to 00 of the posterior distribution given in (111.3)

leads to a well-defined posterior which can be written as

(IV. 1)
 -23-

From (111.8), the corresponding limit applied to Y can be written as

j

where nj = -Y/2j . While a diffuse prior is the working assumption for

the bulk of the empirical analysis, some estimates are provided for

finite values of b to demonstrate the effe~of an upper bound restric-

tion on the model estimates.

The most important choice for the estimations

is the selection of an appropriate method to generate the time series for

the market variance. The derivations in Sections II and III assumed

that a2 et) is observable. Of course, it is not, and therefore, must

be estimated. As discussed in the'~ntroduction," this is not a paper

on either variance estimation or variance forecasting. Hence, a simple

variance estimation model is used. The use of estimated values for the

time series of variances introduces measurement error into the model

.,

estimators. Given the exploratory nature of the paper and the relatively

unsophisticated varian~estimationmodel, no attempt is made to adjust

for these measurement errors. In using the estimation formulas from

Section III, it is assumed that the estimated variances are the true

values of the variances. This is the principal reason why the empirical

results presented here must be treated as "preliminary" and it is also

the reason why no significance tests are attempted.

As discussed in Appendix A, a simple but reasonable estimate for

the monthly variance is the sum of the squares of the daily logarithmic

returns on the market for that month with appropriate adjustments for

weekends and holidays and for the "no-trading" effect which occurs with a
•

-24-

portfolio of stocks. Unfortunately, daily return data for the NYSE Index

is available only from 1962 to 1978. A long time series is essential

for estimating expected returns on stocks and sixteen years of data is

not a long time series. Therefore, to make use of the much lonper monthly

time series, a variance estimator using monthly data was created by

averaging the sum of squares of the monthly logarithmic returns on the

market for the six months just prior to the month being estimated and for

the six months just after that month. That is, the estimate for the
"2
variance in month t, cr (t), is given by

"2
cr (t) = L
{ 6
(In[~(t+k)])2 + L:6 (In[~(t-k)])] } 112. (IV.3)
k=l . k=l

With this variance estimator, all the available market return data except

the first six months of 1926 and the last six months of 1978 can be used

to estimate the models.

Although no explicit c~nsideration is given to measurement errors in

. the variances, some indication of their effect on the model estimates

is prOVided by estimating the models using both the daily return and the·

monthly return estimates of the variance for the period July 1~62 to June

1978.

In Table IV.l, estimates for Model 111 are reported for the two

different variance estimates and for different values of the upperbound

restriction on Y under the assumption that Y is constant over this

l l
sixteen-year period. As might be expected, for a "tight" prior (i.e., b

small) and Y different from the prior expected value of

l
 -25-

YI(YI ~ b!2), the data have little weight in the posterior estimate Y •
I
For this reason, with b
small, the differences in Y for the two
I
different variance estimators are quite small. As b is increased,

the data have greater weight in the estimate of Y and the effect on
I
YI of the different variance estimators also increases. Figure IV.l plots
--- ~ - __ ..
. __~ .. ..

Yl as a function of b using the daily data estimates of the

monthly variances. As shown there,

the differences between Y for b as small as six and Y for the

l l
diffuse prior (b = 00) are rather small. Since no information is

available which might provide a meaningful upperbound on aggregate

relative risk aversion, the effects.of a finite b are analyzed

no further and the diffuse prior assumption (b = 00) is made for the

balance of the paper. That is, only the nonnegativity prior restriction

is imposed upon the Y., j = 1,2,3.

J
From Table IV.l, the effect of the two different variance estimators

on the estimates of Model #1 is moderate with a percentage difference in

. the unrestricted regression estimate Y of about five percent and a

l
percentage difference in the posterior estimate Y of about four
l
percent. Both model parameter estimates were larger for the monthly-

data variance estimates. However, as reported in Table IV.2, the

effect on the estimates of MOdels #2 and #3 is in the opposite direction

and of considerably larger magnitude. As with Model #1, the percentage

differences in the posterior estimates are somewhat smaller than the per-
centage differences in the unrestricted regression estimates for both

Models #2 and #3. However, for all estimates in these latter two models,

the percentage differences are in excess of 30 percent. The effect of the

two variance estimators on the posterior density functions for each of

 Table IV.1

The Effect of Daily Data Versus Monthly Data Estimates for

Variance on Model #1 Estimates For Different Prior Restrictions
July 1962 - June 1978
Least-
Squares
Estimate Model Estimate Y1 For Different Prior Restrictions
2 ")
01 Y1 ~=0.5 b=l b=2 b=3 b=4 b=5 b=6 b=oo ,
MOnthly Data Esti- 0.3482
"-
1.5914 .2597 0.5312 1.0653 1.5215 1.8436 2.0213 2.0931 2.1180
mates of Variance

Daily Data Estimates 0.3733 1.5181 0.2598 0.5312 1.0612 1.5045 1.8054 1.9605 2.0173 2.0341
of Variance

Percentage Dif- -6.72% 4.83% -0.04% 0.0% 0.39% 1.13% 2.12% 3.10% 3.76% 4.12%
ference
 , ;;,
}

, , ,

FIGURE IV.~ TIfE EFFECT of PRIOR IA.rreR-60lUJD RE:>TRICT'OtJS OM MODEL E5TIH~T£1 ~(b)J
us,,.,c; I>AIL'i DATA f~T'HAT'S OF MONTHL'( VARIAtJCE'
.JUL~ l~fD~- 'J'C.UJE let7B

~(b)
3·0

)
=l.S
,

"

,- " Y[oo) ="1. 0:5i1' , '

-- - -----
:,
.,.
I~'O
'
, -- -
-
- 1 - · - as -, -: ~ -: - i---

1\
_ _ _ ' _ . _ _ --..-;._ ---fJ"-- ~ ~_ -l1_=--1€lB1.. . _~_
j i.!
_._.:
. I -- r
, '
, ,

MODEl. * 1: «l-t)- rlt)::: Y~ (f'Ct) i

1
I
' ,1.0'
_... _- .. ~.
.j .

o.s . ---_ .., .... _._ .._.. _.

o r I I I I I ••• I • I I I' b
o O.S 1.0 1.5', -1.D :l.S, 3.0 3.5" ...... 0 1·!, 5.0 "S.S, '.0 '.5,
 Table IV.2

The Effect of Daily Data Versus MOnthly Data Estimates

of Variance on Different Models Estimates
with Nonnegative Restriction Only

July 1962 - June 1978

MOnthly Data Estimates of Variance 0.3482 1.5914 2.1180

Daily Data Estimates of Variance 0.3733 1.5181 2.0341
Percentage Difference -6.72% 4.83% 4.12%

Model #2: ~(t) - r(t) = Y2cr (t)

MOnthly Data Estimates of Variance 192 0.1123 0.1214

Daily Data Estimates of Variance 192 0.1806 0.1818
Percentage Difference 0% -37.82% -33.22%

Model #3: ~(t) - r(t) = Y3

Monthly Data Estimates of Variance 172185 0.0052 0.0053

Daily Data Estimates of Variance 221708 0.0082 0.0083
Percentage Difference -22.34% -36.59% -35.37%
 FIGURE IV.t POSTERIoR.. PRoSABILH'YDISTRIl3UTloN Fori Yi,
. t'loD&L # 1: c( Lt)- riot) lIII '11 a& ") .
. JULY ",% - 3"IJNi '97'
I
•I R

,
R
••
• ,••
•I •l- . . DAILY &~TU1ATE
•I I OF MONTHLY VARI"~C{
.. .
-
I

.. '..,,,,,Y_.
I
I : ,a NOIINEGAT.vE PRIoR.,
:
" .
Re$TRICTloN ONLY
·'0
I
• ' ,i
Vi.'= I.S'111---.• : 2.01411 (b=oo)
i

- .,• '.,
•
•
••
I

.0 •
1-------I-+-.--ti----I------t....---+--==:::;.-::=--.... '11-
0·0 1·0 S.O '.0 7·0
+_
• ,_:- • __ ~ •• _ . _~A'"

,
.... :..._." .. -.- ; ...

. ,
J--:-
·
, I .
,
· I MOm'HL Y EST'MATE
· I . OF MOIJTHI..Y VARIANt!
I .
· I·'
· I . NON tlEGATIV£ PRIOR.
I
I · I .RESTR.ac:.TlcN ON" Y
,. I I . .<.. b=oo)
Yj.=l.seUi ~ ~Yi.: %.1110
I I
I l
I . I
I I
I . I·
I

o--+-~t.HO--3-4.-0--41 ....-0-·--s+.0--=:::'+='"0::::::=='-'-+7./ 1
·0 . o,&...o-.--.....

i
.. _ - ,--~. -. _ ,., -.. .~. - .." .-.. '--""".,-, _. ..
 F.GIlRI IV. '5 PO'TER,olt PROSABILITY DiSTRIBUTION FOR Va.
=
MODIL # 2. : o((i)-t-tt) Va. O""lt)
Jw.y 11t-1- "JUliE 'f7'

DA'LY £5T' MATE

. . OF HOttlTHL Y V/fRIANCE

. NOtJNE~"T'v£ PRIoR.
R&STR'C.T'OAI ONLY
(b- 00) .

0-10 0·$0

MONTH..Y eSTIMATE
1~o
OF NoN'httY v"'lAtleS

NOH NI4ItT,V£ PRIoR.

RISTRIC.lION ONLY
. . .c"= 00)
A , _

V.... O.U2!.~ '.IYI.=O.llI1

I ,
I I
. .... . I I
I I

0·30 0·10
 POSTER'oR. PROBAalUTY DI$TIU 8L1T'O~ FoR. Y.
MoO'1. # S: ot(t)-rLt):= 'is
___ _ . __ _ . 3"IlLY ., '1- :r...". '1"
_•... __ ._._,_ .... _._. __. .. _. _ _ ._... _.. .. •._._. r o ' __ ' •. _ . - • • • • •".

.
~
,
_
~
.
-
~
_
.
_
-
,

,so-o
__ _DAII.'t e,r'"ATE
_ _____. __ OF "0"'11a. y ",,,,.AIICE

'00-0 I ___.. NOtI NIGAT.VI PRIoR. . j;

I
I . __ . R&STIUt.T,olJ OHI.Y
I _ .(b :.0)
I
I .
I
I
A I

Yr·oot~Ys=·OO8.t

-0·0 o -ooSo O· 0100 0.0'500-0:&00 0-02S0

- ~oo-o

- ISO'O

"'OF' TMo-,THLY
NI.'1 VAR,lUIcG £ST'HAT£

_. ... ...No,UIEGATIVE PRioR

_ RiSTRIC.TION ONLY
_.. _.... (b= 00)

YS=.OOS%"'tVYS: .00SI
-0·0
o-ooso O· 010..0 0-01 SO O-OZOO
 -26-

the three models are illustrated in Figures IV.2, IV.3, and IV.4.

While these brief comparisons ,cannot be considered an analysis of the

effects of measurement error in the variance estimates, they do serve

as a warning against attaching great significance to the point estimates

of the models.

From Table IV.2, it appears that for this period, the prior nonnega-

tivity restriction is important only for Model #1 where for the same

variance estimate, the percentage difference between Y and Y is

l l
approximately 25 percent. The differences between the posterior

estimate and the unrestricted regression estimate for Model #2 and Model

#3 are negligible. This result is further illustrated by inspection of

the shapes and domains of the posterior density functions as

plotted in Figures IV.2, IV.3, and IV.4.

To further investigate the importance of the prior nonnegativity

restriction, the differences between the pnoterior and unrestricted

,--
regression estimates are examined for fifty-tvlO years of data from

July 1926 to June 1978. These estimates are presented in Table IV.3

for both T = 52 years and T = 26 years. As inspection of this table

immediately reveals, the percentage differences between Y. and Y.

J J
are negligible for all three models with T = 52, and for Models #2 and

#3 with T = 26. For Model #1 with T = 26, the differences are small

with an average about half of that found in the previous analysis from

1962-1978. As before, the posterior density functions for each of the

models with T = 52 are plotted in Figures IV.5, IV.6, and IV.7.

By the assumption that the Y. , j = 1,2,3, are constant over such a

J
 Table IV.3

Different Model Estimates for 52~Year and 26-Year Time Intervals

July 1926 - June 1978

52-Year Intervals 26-Year Intervals

7/26-6/78 7/26-6/52 7/52-6/78 Average

2
Model //1 : a.(t) - r(t) =Y CJ (t)
1

ni 2.16246 1.6617 0.5007 1.0812

" 1.8932 1.5112 3.1608 2.3360

Y
1

Y 1.8988 1.5588 3.2076 2.3832

1

Percentage Difference -0.30% -3.05% -1.46% -2.26%

Model /12: a.(t) - r(t) = Y2CJ(t)

0.12 624 312 312 312

"
Y .1867 .2012 .1723 .1867
2
,/
Y .1867 .2012 .1725 .1869
2

Percentage Difference 0.0% 0.0% -0.16% -0.08%

Model /13 : a.(t) - r(t) = Y

3

rl3 423624 144884 278740 211812

"
Y .0082 .0109 .0068 .0089
3
-
Y3 .0082 .0109 .0068 .0089

Percentage Difference 0.0% 0.0% 0.0% 0.0%

 ... FIGIA.RE Iv.r POSTERIOfl PROBABILITY DlsTRlauT'o~ Fort Ys.
"ODIiL #' ~: at (i)- r(t). y~ (fIt'll
.............. .;rI&LY 1121. - ;JUliE .17'
. .... . .. --.. ..
•

.. ....•t.o

I
I
I
I
•I
·········10 I
•I
. NOH IIIGAT'VE PRIOR.
I
I . "STRIC.T'ON ONLY
•I ..._.._...( b. CO)
/ ..
. I ..
II> . I
\:I.I'1t,-~1'11.1"r
I
.... .. ·0
Ys.
0'0 3·0 1-0 s·o "·0

~ __ ... -"._ ._ ._ ....__•.. __ ._ .. __ ..... _L.".

 POSTERIOR. PROBABILITY 1),sTR,8KTloAl FOR VI
MOl)£ L # 2: Q(t~)- ...tt) = '(I. cr(.)
.. __ __ . _ ~ItL Y 111f. - ~UHE 7, J'

·S'·O
NoN NEGAT.VE PR/ott
RISTRIC.TION OIlL Y
t
(b= (0)
I
I
I
A ,_

'1t:""~I.,...YI.:.llc."7
.....-_IIIIt:;..~ .......__ ~~-._ .....__ ~.....- __......_ YI.
. 0·0 0 · '0 0 · ao 0 · SO tJ • 410 0 · 50 0· '0

. "-_ ..- _
..•......•. ' .~..:._, - ... -- .. '. _.. _. -', -, ...... - -- .-- .. --- .. _ .._-- .. -._.". _. _. ".
 FI6UR£ IV.1 PoSrlRloR. PRolABIUTY DISTR'.WT,oN FoR. Ys
'NODaL # I: ottt)- rlt): Y.
_._. _. ... . . JULy ",,,- 3'Ur4E Ict."

- -t.o·o

. _ NON Ntt;AT,vE PI,O..

- .100·0 _.__ RISTR.C.T,ON Or4t Y
_. __ (bcoo)

.- 0-0
.
l=.ootl
I ...
-
y~.ooal
..,s
0·0 o· 0050 0·0100 0'0,50 0·0100 o·oaso

---"'--'--'- -'-- ._.. - _.~ .---.. _._..- -~ ...•.. -. --~-- -- .-.- - ~. __ .- ".•--- .- •.. -.-._ .." .'.f"-- .-. -
~ -•• -'" -. , '." ..

- .'

I
•_ ......•.....•.._ •. _.__ , ...•....... ~ _~_. ._._~_ . _ . __ ."._. ._0_ .• ~_.,., 1.
 -27-

long time period, the number of observations N is quite large (624

for T = 52 and 312 for T == 26). Given the previously-demonstrated

,..
asymptotic convergence of Y. +Y. for large N, these findings were
J J
not entirely unexpected. However, if shorter time intervals over which

Y is assumed to be constant are chosen, then the differences between

j
Y. and Y. are not negligible.
J J
In Table IV.4, the different model estimates are presented for

T = 13 years (with N = 156). The average percentage difference

'j:

between Y. and Y. for the four l3-year time periods ranged from
J J
a high of 28 percent for Model /11 to a low of 6 percent for Model f!3

with Model #2 in the middle at a 12 percent difference. However, the

percentage differences for each of the time periods are more important

than the average since by hypothesis, the Y. can only be estimated

J
using 13 years of data.

In the 1965-1978 {ieriod, the percentage differences between the

posterior estimate and the unrestricted regression estimate are

substantial for all three models. This was a period with a number of

large negative realized excess returns on the market, and this is

precisely the tYpe of period in which the prior nonnegativity restriction

can be expected to be important. The periods 1939-1952 and 1952-1965

did not have these large negative realized excess returns and corres-

pondingly, the nonnegativity restriction was (expost) unimportant.

The period 1926-1939 appears to be different from the other three

in that the effect of the nonnegativity restriction is quite larg~ for

Model /11; small for Model //2; and negligible for Model //3. However,
 Table IV.4

Different MOdel Estimates for 13-Year Time Intervals

July 1926 - June 1978

7/26-6/39 7/39-6/52 7/52-6/657/65-6/78 Average

2
Model 111: a(t) - r(t) =Y 0" (t)
1
2
0. 1.3344 0.3273 0.1936 0.3072 0.5406
1
"
Y1 0.6281 5.1114 7.5772 0.3777 3.4236

Y1 0.9747 5.1211 7.5807 1.5858 3.8156

Percentage Difference -35.56% -0.19% -0.05% -76.18% -28.00%

Model /12: a(t) - r(t) = Y 0"(t)

2
2
0. 156 156 156 156 156
2
"
Y2 .1569 .2454 .2982 .0464 .1867

-
Y2 .• 1617 .2457 .2983 .0840 .1974

Percentage Difference -2.97% -0.12% -0.03% -44.76% -11. 97%

Model 113: a(t) - r(t) = Y

3
2
0. 44439 100445 164110 114630 105906
3
" .0146·
Y .0092 .0096 .0029 .0091
3
Y .0146 .0092 .0096 .0038 .0093
3

Percentage Difference 0.0% 0.0% 0.0% -23.68%· -5.92%

 -Z8-

the results from this period are consistent with the others. This was

a period of both large positive and negative realized excess returns

with both large changes in variance and large variances especially in

the early 1930's when the market had a large negative average excess

return. From the regression estimators, (II.Z), Y

" has in its
l
numerator the unweighted average of the (logarithmic) realized excess

returns. Y has in its numerator the weighted average of these

Z
excess returns where the weights are such that each excess return is

"deflated" by that month's estimate of the standard deviation. That is,

unlike "
Y in which each observed excess return has the same weight,
l
Y puts more weight on observed excess returns which occur in lower-
Z
than-average-standard-deviation months and less weight on those that

occur on higher-than-average-standard-deviation months. Inspection

of the regression estimator for Hodel 113 will show that the weighting

of the realized excess returns is similar to that of Y except the

Z
effect is more pronounced because each month's return is divided

. by that month's variance. Hence, in a period such as the early

1930's when, expost, large negative excess returns occur in months

where the variance is also quite large, the differences between Y.

J
and Y. will be largest in Model #1 and smallest in Model #3. Of
J
course, just the opposite effect will occur in periods when, expost,

either large negative excess returns occur in months when the variance

is small, or more likely, large positive excess returns occur in months

when the variance is large and a number of negative excess returns

occur in months when the variance is small.

 -29.-

To provide further evidence in support of this explanation and to

further underscore the importance of the nonnegativity constraint,

especially as T becomes smaller, Tables IV.5.l, IV.5.2, and IV.5.3

provide the estimates for all three models for T =4 years (N = 48).
In the 1930-1934 period, the regression estimates were negative for all

three models with the largest percentage difference between Y. and

J
Y occurring for Model #1 and the smallest for Model #3. In the 1938
j
-1942 period, the regression estimates for Model #2 and Model #3 were

negative, and the ranking of the models by percentage differences between Y.

J
and Yj was reversed from that of the 1930-1934 period. In the 1966-1970

period, the regression estimates for Model #1 and Model #2 are

negative with the same model rankings as in the 1930-1934 period.

While the periods in which the regression estimates are negative

demonstrate the necessity for the nonnegativity restriction most

dramatically, it is not necessary that the estimatesbe negative to

have large percentage differences between Y. and Y.. Two cases

J J
in point 3re the 1970-1974 and 1974-1978 periods. As expected,

when T is further reduced from four years to one year, the effect

of the nonnepati~ty.restrictionis even more. pronounced. The

summary statistics for this case are presented in Table IV.6.

As a final illustration of the necessity for including the

nonnegativity restriction, the estimates of Yj and Y

j
(j = 1,2,3)
using the monthly...data variance estimator for the period 1962-1978

are compared with the corresponding estimates for the period 1965-1978.
 Table IV.5.1

MOdel H1-Estimates for 4-Year Time Intervals

July 1926 - June 1978

(l,(t) - r(t) = Y1cr 2 (t)

,.,
Percentage
r"z1 Y
1
Y
1 Difference

7/26 - 6/30 .1785 2.4778 3.1184 -20.5%

7/30 - 6/34 .8365 -0.1097 0.8337 -113.2%
7/34 - 6/38 .2276· 1.9389 2.6017 -25.5%
7/38 - 6/42 .2226 0.2156 1.7721 -87.8%
7/42 - 6/46 .0860 12.6511 12.6525 0.0%
7/46 - 6/50 •0787 1.9159 . 3.6625 -47.7%
7/50 - 6/54 .0553 12.2185 12.2459 -0.2%
7/54- 6/58 .0685 . 7.6509 7.8610 -2.7%
7/58 - 6/62 .0607 4.3146 5.3906 ""'20.0%
7/62 - 6/66 .0507 9.0518 9.2786 -2.4%
7/66 - 6/70 .0863 -2.2060 2.0534 -207.4%
7/70 - 6/74 .0909 0.9986 3.0443 -67.2%
7/74- 6/78 .1204 1.6183 2.9964 -46.0%

Average .1664 4.0566 5.1932 -49.3%

 Table IV.5.2

MOdel H2-Estimates for 4-Year Time Intervals

July 1926 - June 1978

aCt) - ret) =Y2a(t)

A
2 Percentage
0 Y2 Y2'
2 Difference
7/26 - 6/30 48 .2658 .2768 -4.0%
7/30 - 6/34 48 -.0084 .1122 -107.5%
7/34 - 6/38 48 .2549 .2675 -4.7%
7/38 - 6/42 48 -.1288 .0790 -263.0%
7/42 - 6/46 48 .5509 .5510 0.0%
7/46 - 6/50 48 .1206 .1715 -29.7%
7/50 - 6/54 48 .4109 .4119 -0.2%
7/54 - 6/58 48 .2954 .3027 -2.4%
7/58 - 6/62 48 . .2171 .2370 -8.4%
7/62 - 6/66 48 .3293 .3336 -1.3%
7/66 - 6/70 48 -.0355 .1032 -134.4%
7/70 - 6/74 48 .0653 .1424 -54.1%
7/74 - 6/78 48 .0901 .1547 -41.8%

Average 48 .1867 .2418 -50.1%

 Table IV.5.3

MOdel #3~Estimates for 4-Year Time Intervals.

July 1926 ~ June 1978

a(t) - r(t) = Y3
x
Percentage
~i3 Y
1
Y
3 Difference

7/26 - 6/30 22344 .0165 .0167 -1.2%

7/30 - 6/34 4091 -.0015 .0119 -112.6%
7/34 - 6/38 16124 .0183 .0185 -1.1%
7/38 - 6/42 19633 -.0149 .0026 -673.1%
7/42 - 6/46 29693 .0222 .0222 0.0%
7/46 - 6/50 33517 .0062 .0075 -17.3%
7/50 - 6/54 47626 .0126 .0126 0.0%
7/54 - 6/58 35062 .0111 .0113 -1.8%
7/58 - 6/62 43347 .0084 .0088 -4.5%
7/62 - 6/66 73342 .0088 .0089 -1.1%
7/66 - 6/70 30920 .0014 .0051 -72.5%
7/70 - 6/74 35864 .0031 .0056 -44.6%
7/74 - 6/78 32059 .0029 .0057 -49.1%

Average 32586 .0073 .0106 -75.3%

 Table IV.6

Summary Statistics of Different MOdels Estimates

for I-Year Time Intervals

July 1926 - June 1978

MOn~h1y Estimates for 52 I-Year Intervals

Standard
Average Deviation High Low

MOdel 111: a.(t) - r(t) = Y1a 2 (t)

,.
Y 4.7982 8.4217 26.2476 -9.5025
1

Y 8.5001 6.0049 26.3422 0.7471

1

MOdel #2: a.(t) - r(t) = Y2a(t)

0.1867 0.3791 0.8987 -0.6214

0.3719 0.2086 0.8996 0.1029

MOdel 113: a,(t) - r(t) = Y

3
,.
Y 0.0061 0.0316 0.1322 -0.1119
3

Y 0.0181 0.0173 0.1323 0.0040

3
 -30-

Since the variance estimates and return data axe identical for the l3-year

overlapping period 1965-1978, the differences between the estimates

presented in Table IV.2 and those presented in Table IV.4 reflect the

effect of a change from a l6-y.ear to a l3-year observation period. The

three-year period 1962-1965 eliminated by this change was one in which

the realized excess returns on the market were mostly positive .and the

variances were relatively low.

For Model #1, the effect of this change on the posterior estimate

Y is a 25.1 percent decline. While this was substantial, the effect

l
~
on the regression estimate was much greater with a decline in y of
1
76.3 percent. The effect on the other model estimates is similar.

For Model #2, the posterior estimate Y2 changes by 30.8 percent

A

with a corresponding change in Y of 58.7 percent. For Model #3,

2
the change in Y is 28.3 percent and the change in Y is 44.2
3 3
percent.

The substantial percentage change in both the Y and Y.

j J
. estimates from a relatively small change in the observation period

illustrates the general difficulty in accurately estimating the

parameters in an expected return model and underscores the importance

of using as long a historical time series as is available. However,

very long time series are not always available, and even when they

are, it may not be reasonable to assume that the parameters to be

estimated were stationary over that long a period. Therefore, given

the relative stability of the Y. ·estimator by comparison with Y.,

J J
it appears that the nonnegativity restriction should be incorporated
 -31-

in the specification of any such expected return model.

Having analyzed the empirical estimates of the Y., we now examine

J
the properties of the expected excess returns on the market implied

by each of these models. For this purpose, it is ,!ssumed that the

Y ., (j
j
= 1,2,3), were constant over the entire period 1926-1978, and

therefore, T equals 52 years. Of course, this assumption is

certainly open to question. However, given the much-discussed problems

with the variance estimators and the exploratory spirit with which

this paper is presented, further refinements as to the best estimate

of T are not warranted here. Moreover, as discussed in Section II,

the current state-of-the-art model implicitly makes this assumption

by using as its estimate of the expected excess return on the market,

the sample average of realized excess returns over the longest data

period available.

Using the estimated Y. and the time series of estimates for the
J
market variances, monthly time series of the expected excess return on

- the market were generated for each of the three models over the 624

months from July 1926 to June 1978. As shown in Figures IV.5, IV.6,
and IV.7, with T equal to 52 years, the posterior density functions

for all three models are virtually symmetric and the differences between Y.
J
and Y. are negligible. Hence, for T = 52 years, the monthly time series
J
of expected excess returns using the unrestricted regression estimates

would be identical to those presented here.

The. summary statistics for thesemon~hly time series are reported

in Table IV.7 and they include the sample average, standard deviation
 -32-

and the highest and lowest values. Of course, the expected excess

r~turn estimate for Model 113 is simply a constant. In Table IV.7,

the same summary statistics are presented for the realized excess

returns on the market and for the realized returns on the riskless

asset.

Inspection of Table IV.7 shows that the average of the expected

excess returns varies considerably across the three models. The

"Constant-Preferences" Model III is the lowest wUhan average of 0.665

percent per month or, expressed as an annualized excess return, 8.28

percent per year. The "Cons tant-Price-'o f-Risk" Model 112 is the

highest with an annualized excess return average of 12.04 percent per

year. The "Constant-Expected-Excess-Return" Model 1/3 is almost exactly

midway between the other two models with an annualized average of

10.36%. The sample average of the realized excess returns on the market

was 0.655 per.:'-ent per month, or, annualized, 8.15 percent per year •.

This sample average is also the point estimate for the expected excess

. return on the market according to the state-of-the-art model.

Even with these large differences in the average estimates, it is

unlikely that any of these models could be rejected by the realized

return data. The variance of the unanticipated part of the returns

on the market is much larger than the variance of the change in expected

return. That is, the realized returns are a very "noisy" series for

detecting differences among models of expected return.

In examining the average excess returns in Table IV.7, one might

be tempted to conclude that Model III "looks" a little better because its
 Table IV.7

Summary Statistics of Model Estimates for the Monthly

Expected Excess Return on the Market and Sample Monthly
Realized Returns for the Market and U.S. Treasury Bills
July 1926 - June 1978

Monthly Expected Excess Return on the Market

Standard
Average Deviation 1!!.8!!. Low

Model /11: aCt) - ret) = Y1C1 2 (t) 0.665% 1.032% 7.161% 0.048%

Model #2: aCt) - ret) = Y2C1(t) 0.952% 0.570% 3.628% 0.297%

Model #3: aCt) - ret) = Y3 0.825%

Monthly Realized Returns

Standard
Average Deviation High Low
U.S. Treasury Bills: ret) 0.207% 0.184% 0.810% -0.240%

NYSE Excess Returns: 0.655% 5.881% 38.408% -29.137%

l\r(t)/[1 + ret)] - 1
 -33-

average is so close to the sample average6f realized excess returns.

However, as inspection of (11.2.1) makes clear, the regression estimator

Y is such that this must always be the case when the variance estimator
l
is of the type used here. This observation brings up an important issue

with respect to estimates based upon the state-of-the-art model.

If the strict formulation of that model is that the expected

excess return on the market is a constant or at least, stationary over

time, then the least-squares estimate of that constant is given by

Y in MOdel #3. However, from Table IV.7,the annualized difference

3
between Y and the sample realized return average is 221 basis points.
3
This difference is quite large when considered in the context of

portfolio selection and corporate finance applications. The reason for

the difference is that the sample average of realized returns is only

a least-squares estimate if the variance of returns over the

period is con.::~ t.ant • If the variance is no t cons tan t, and it isn't,

then the estimator should be adjusted for heteroscadasticity in the

. "error" terms. This is exactly what the estimator Y does.

3
Of course, the sample average of realized returns is a consistent

estimator and the measurement error problem in the variance estimates

rule out formal statistical comparison. However, the large difference

reported here should provide a warning against neglecting the effects

of changing variance in such estimations and simply relying upon

"consistency" even when the observation period is as long as 52 years.

As mentioned, the sample average of the realized returns will

provide an efficient estimate of the average expected excess return if

 -34-

Model 111 is the correct specification. However, even if that is the

belief, then for capital market and corporate finance applications,

Yl times the current variance will provide a better estimate of the

current expected excess return than the state-of-the-art model because

it takes into account the current level of risk associated with the market.

A similar argument applies to using the ratio of the sample

average of the realized excess returns to the sample standard deviation

for estimating the Price of Risk under the hypothesis that it is

constant, or at least, stationary over time. From Table IV.7, using the

realized return statistics, the estimate of the Price of Risk is 0.114

per month whereas the least-squares estimate Y which takes into

2
account the changing variance rates is 0.1867 per month. Again, this

difference is quite large.

To further underscore the importance of taking into account the

change in the variance rate when estimating the expected return on the

market, we close this section with a brief examination of the time series

of market variance estimates. The average monthly variance rates for the

market returns are presented in Table IV.8 for the thirteen successive

four-year periods from July 1926 to June 1978. Over the entire 52-year

period, the average annual standard deviation of the market return was

20.4 percent. However, as is clearly demonstrated in Table IV.8, the

variance rate can change by'a substantial amount from one four-year

period to another, and it is significantly different from this average

in many of the four-year periods.

 Table IV.8

Successive Four-Year Average Monthly Variance Estimates for

The Return on the Market

July 1926 - June 1978

Average Monthly Variance* Percentage Change

Dates a2 (t) From Previous Period

7/26 - 6/30 .003719

7/30 - 6/34 .017427 368.59%
7/34 - 6/38 .004742 -72.79%
7/38 - 6/42 .004638 -2.19%
7/42 - 6/46 .001792 -61.36%
7/46 - 6/50 .001640 -8.48%
7/50 - 6/54 .001152 -29.76%
7/54 - 6/58 .001427 23.87%
7/58 - 6/62 .001265 -11.35%
7/62 - 6/66 .001056 -16.52%
7/66 - 6/70 .001798 70.27%
7/70 - 6/74 .001894 5.34%
7/74 - 6/78. .002508 32.42%

Average .003467

*Using formula (IV.3) for the variance estimator~

 -35-

It has frequently been reported that the market was considerably

more volatile in the pre-World War II period than it has been in the

post-war period. That observation is confirmed here with an average

annual standard deviation of 27.9 percent for the period July 1926 to

June 1946 versus 13.8 percent for the period July 1946 to June 1978.

However, a significant part of this difference is explained by the

extraordinarily large variance rates in the 1930-1934 period. Thus,

if this period is excluded, then the average annual standard deviation

for the other twelve four-year periods is 16.6 percent.

Because the state-of-the-art model assumes a constant variance

rate, the large differences in variance rates among the various subperiods

causes this model's estimates to be quite sensitive to the time period

of history used. So, for example, if the 1930-1934 is excluded, then

th~ estimated Market Price of Risk based upon the other forty-eight

years of data changes by 33 percent for the state-of-the-art model

estimator. However, this same exclusion caases Model #2's estimate, Y2'

. to change by only 8 percent.

 -36-

V. Conclusion

In this exploratory investigation, we have established two substantive

results: First, whether or not one agrees with the specific way in

which it was incorporated here, it has been shown that in estimating

models of the expected return on the market, the nonnegativity restriction

on the expected excess return should be explicitly included as part of

the specification. Second, because the variance of the market return

changes significantly over time, estimators which use realized return

time series should be adjusted for heteroscadasticity. As suggested by

the empirical results presented here, estimators based upon the assumption

of a constant variance rate, although consistent, can produce substan-

tially different estimates than the proper least-squares estimator even

when the time series are as long as fifty years. As demonstrated by the

analysis of :Hode1 113, these conclusions apply even i f the model specifi-

cation is such that the expected excess return does not depend upon

the level of market risk.

There are at least three directions in which further research along

these lines may prove fruitful. First, because the realized return data

provide "noisy" estimates of expected return, it may be possible to

improve the model estimates by using additional nonmarket data. Examples

of such other data are the surveys of investor holdings as used in Blume

and Friend (1975); the surveys of investor expectations as used in Malkiel

and Cragg (1979); and corporate earnings and other accounting data as

used in Myers and Pogue (1979). Because these types of data are not

available with the regularity and completeness of market return data, it

may be more appropriate to include them through a prior distribution

 -37-

rather than as simply additional variables in a standard time series

regression analysis. If a prior distribution is to be used to incorporate both

these data and the nonnegativity restriction, then the sensitivity of

the model estimates to the particular distribution chosen warrants careful

study.

A second direction for further study is the length of time over which

it is assumed that the Y. can be treated as essentially constant (i.e., T).

J
In the analysis presented here, the estimates of Y. for different T
J
only used the data for the specific subperiod. So, for example, the Y.
J
for the period 1930-1934 was computed using only the observed returns

for 1930-1934. Clearly, better estimates could be obtained by including

the pre-1930 observations as well. Therefore, for a given T, the

estimates will be improved by developing a procedure for revising the

prior distribution using past estimates of Y.•

J
The third and most important direction is to develop accurate

variance estimation monels. As previously discussed, such models have

. applications far broader than simply estimating expected returns. Such

models should benefit from inclusion of both option price data and

accounting data in addition to the past time series of market returns.

Perhaps other market data such as trading volume may improve the estimates

as well.

While there are obviously many problems to be overcome in both

the estimation and testing of expected. return models, it is hoped that

this paper will stimulate further research effort and with it, some

solutions to this important problem.

 Fl

FOOTNOTES

*Aid from the National Bureau of Economic Research and the National
Science Foundation is gratefully acknowledged. My thanks to F. Black
and J. Cox for many helpful discussions and to R. tl~nriksson for
scientific assistance.

1. See Sharpe [1964], Lintner [1965], and Mossin [1966]. For an

excellent survey article on the Capital Asset Pricing Model, see
Jensen [1972].

2. See Jensen [1972]; Black, Jensen, and Scholes [1972]; Fama and McBeth
[1974]; and Frien.d and Blume [1970].

3. See Roll [1977].

4. See Breeden [1979]; Cox, Ingersoll, and Ross (forthcoming); Ross

[1976]; Merton [1973] and [1980].

5. By "dominant factor," we do not mean that most of the variation in an

individual stock's realized returns can be "explained" by the variation
in the market's return. Rather, we mean that among the systematic
risk factors that influence an individual stock's equilibrium expected
return, the market risk of that stock will have the largest influence
on its' expected return.

6. See Fama and Schwert [1977].

7. Sufficient conditions for this propositIon to obtain is that all

investors are strictly risk-averse expected utility maximizers. For
a proof of the proposition, see Merton [1980; Proposition IV.6].

8. See Ibbotson and Sinquefield [1977]. IIi Ibbotson and Sinquefield

[1979, p. 36], they express the view that "The equity risk premium
has historically followed a random walk centered on an arithmetic mean
of 8.7 percent, or 6.2 percent compounded annually."

9. It is shown in Rothschild and Stiglitz [1970] that the demand for a

risky asset in an optimal portfolio which combines this asset with the
riskless asset, need not be a decreasing function of the risk of that
asset.. Hence, it is possible that an increase in the riskiness of the
market will not require a corresponding increase in its equilibrium
expected return. For further discussion of this point,. see Merton
[1980].
 F2

10. For a development of the continuous-time model with diffusion-type

stochastic processes, see Merton [1971;1978;1980]. As is discussed
at length in these papers, the assumptions of continuous trading anu
diffusion-type stochastic processes justify the use of variance as a
sufficient statistic for risk without the objectionable assumptions
of either quadratic utility or normally-distributed stock returns.

11. In the intertemporal model presented in Merton [1973], (11.6) will be

a close approximation to the equilibrium relationship if either
k
rac /ax.I"« ack/aw, j = l, ••• ,m; k = 1, ••• ,1(, or the variance of
J
the change in W is much larger than the variance of the change in
k
x., j = 1, ••• ,m where C = Ck(W,X,t) is the optimal cons~Ption
J
function of the investor k; W is the wealth of investor k; and
(Xl' ••• '~) are the m state variables (in addition to wealth and
time) required to describe the evolution of ~he economic system.

12. The reader is reminded that if In{E[~(t)/R(t)]} = [a(t)-r(t)]h,

then E{ln[~(t)/R(t)]} = [a(t)-r(t)-o2(t)/2]h.
 Al

Appendix ~

Estimating the Variance Rate From Time

Series of Realized Returns

In the Introduction, it was claimed that the variance of returns can

be estimated far more accurately from the available time series of

realized returns than can the expected return. We now show that this

claim is correct provided that market returns can be described by a

diffusion-type stochastic process as in (11.5) and that the mean and

variance of these returns are .slowly-varying functions of time.

As discussed in detail in Section I I of the text, under the hypo-

thesis that the mean and variance are slowly-varying functions of time,

the true process for market returns can be approximated by assuming

2
that and (J are constants over (nonoverlapping) time intervals

of length h where ~ is the expected logarithmic rate of return on

2
the market per unit time and (J is the variance per uni t time.

Suppose that the real·ized return on the market can be observed over time

intervals of length 6. where 6.« h. Then n:: h/6. is the number of

observations of realized returns over a time interval of length h. So,

for example, if h equals 1 month and 6. equals 1 day, then n equals

30 (neglecting weekends and holidays). Let Xk denote the logarithmic

th
return on the market over the k observation interval of length 6.

during a typical period of length h for k = 1,2, ••• ,n. From (11.5),

Xk can be written as

, k = l,2, ••• ,n , (A.l)

 A2

where the {~}, k = l, ••• ,n are independent and identically distributed

standard normal random variables.

From (A.l), the estimator for the expected logarithmic return,

- [ ~~] Ih ,
'"
II will have the properties that

'"
E[ll] = II (A.2)

and
'"
Var[ll] = (i/h (A.3)

Note that the accuracy of the estimator as measured by Var[ll] depends

only upon the total length of the observation period hand not upon

the number of observations n. That is, nothing is gained in terms of

accuracy of the expected return estimate by choosing finer observation

intervals for the returns and thereby, increasing the number of observa-

tions n for a fixed value of h.

Consider the following estimator for the variance rate:

~2 _ [~~] Ih. From (A.l), this es timator will have the properties

that

E(~2) = cr 2 + 112~ (A.4)

= cr2 + 112h/n

and
"2 = 2cr4 In + 411 2h/n 2
Var(cr ) (A.5)

Because the estimator for cr

2
was not taken around the sample mean
."ll,
"'2
cr .is biased as shown in (A.4). However, for large n, the difference

between the sample second central and noncentral moments is trivial.

 A3

2
For example, typical values for ~ and cr in annual units would be

~.10 and 0.04, respectively. For daily observations, hln will equa1_

approximately 1/360. Therefore, substituting into (A.4), we have that

E(~2) will equal 0.0400277 when the expectation of an unbiased estimator

is 0.04. Even for monthly observations with hln equal to 12,

E(;2) = 0.0408333 and the bias of this estimator is still trivial.

The advantage of this estimator is that the variance can be

estimated without knowing or even having an estimate of the mean ~. It

also, of course, saves one degree of freedom. Thus, for stock return

data and observation intervals of a month or less, the bias from a non-

central estimator of the variance can be neglected.

More important than the issue of bias is the accuracy ?f the esti-
"2
mator. As inspection of (A.s) quickly reveals, Var(cr ) does depend

upon the number of observations n for a fixed h, and indeed, to order

lin, it depends only upon the number of observations. Thus, unlike

the accuracy of the expected return estimator, by choosing finer observa-

tion intervals ~, the accuracy of the variance estimator can be improved

for a fixed value of h.

To further emphasize the point, consider the extreme case where both

the mean and variance are constant for all time. The accuracy of the

expected value estimator will depend only on the total length of

calendar time for which return data are available (e.g., 52 years).

However, the accuracy of the variance estimator will depend critically on

whether these data are available annually, quarterly, monthly, or daily.

The standard deviation of the variance estimate using annual data will
 A4

be approximately ninteen times larger than the standard deviation of

the estimate using daily data over the same calendar period. Since

neither the mean nor the variance are constant for anything like this

length of calendar time, the practical advantage of the variance

estimator's accuracy depending upon n rather than h is that a

reasonably accurate estimate of the variance rate can be obtained using

daily data while the estimates for expected return taken directly from

the sample will be subject to so much error as to be almost useless.

Additional discussion on these points can be found in Merton [1976,

p. 336-339].

In the theoretical limit of continuous observation, n goes to

infinity for any finite h, and therefore, the variance rate could in

principle be estimated without error for any finite interval. However,

in practice, the choice of an ever-shorter observation interval

introduces another type of error which will "swamp" the benefit of a

shorter time interval long before the continuous limit is reached.

This error is caused by not knowing the exact length of time between

successive trades. For example, suppose that the closing price of a

stock as reported in the newspaper was not the result of trade at 4:00

PM, but rather the result of a (last) trade which occurred at 3:00 PM.

If ~ is one week and if the last trade the previous week did occur

at 4:00 PM, then the observed price change occurred over a 167 hour

interval and not a 168 hour interval as assumed. While this actual

shorter time interval will cause an underestimate of the variance rate,

the magnitude of the error is only 0.6 percent. However, suppose that

~ is the six-hour interval from the 10:00 AM opening to the same day
 AS

4:00 PM closing. Then, even if the first trade occurs at 10:00 AM,

the actual interval for the observed price change is five hours and not

six, and the magnitude of the error in the variance rate would be 16.7

percent. Of course, if the first trade actually took place at 11:00 AM,

then the error would be 33.3 percent. Thus, the "true" time interval

between trades is a random variable.

For actively-traded stocks and a ~ of the order of a week or more,

this error will generally have a negligible effect on the variance

estimates. However, for daily observations, it can have a nonnegligible

effect for individual stocks, and it will definitely be important for a

large portfolio of stocks such as the NYSE index. It is more important

for a large portfolio because it is likely that a significant number of

the stocks in the portfolio will not have their last trade near the

closing time. Since the closing value of the index is computed using

the last traded prices and since stocks tend to be contemporaneously

positively correlated with one another, th~ observed daily changes in

the index will exhibit positive serial correlation. This positive

correlation is not "real" in the sense that one could make money trading

in the individual stocks contained in the index because at 4:00 PM,

trades could not have been executed at these last (and earlier) prices.

However, if no adjustment is made for .this "nontrading" effect, then

the sum of squared daily logarithmic changes in the index 'will produce

a significandy biased-low estimate of the variance rate.

A method for correcting for this problem is as follows: Let ~

denote the observed change in the index as contrasted with ~ in (A.I)

 A6

which is the true change. Since it is not known for how many days this

"non-trading" effect lingers, it was simply assumed that after three

days, there is no effect. A model specification which captures this

non-trading effect is given by

'" ..
~ = ~~ + aV'K [oOE:k + 01E:k- l + 02E:k-2 + 03E:k _ 3 ] (A.6)

where 0 ~ OJ ~ 1; j = 0,1,2,3 and 0·o = 1 - ° -0 - °

1 2 3· From (A.6),

we have that

(A.7)

Comparing (A.7) with (A.4) confirms that the unadjusted estimator is

biased low for the posited restrictions on the 0 .• If the were

J
known, then (neglecting the ~2l\ term) there is a simple adjustment:
Namely, divide the observed sum of squares by
[tJ=O
o~J . Of course,

they are not known, and therefore, must be estimated.

In Section IV, daily data are used in one of the variance estimators

for the period July 1962 to June 1978. To adjust the estimates for

serial correlation, the OJ were estimated using nonlinear procedures

on the following equation:

(A.8)

From (A.6) and (A.8), we have that

(A.9)
 A7

(A.9 Cont 'd.)

In the sample period, A = .0002 with standard error .0001 and

B = .2106 with standard error .0114. The corresponding values for OJ

from (A.9) are 00 = .734; 01 = .221; 02 = .045; 03 = .010 •

. 3 2
~ OJ = .5854, and this was the number used to adjust for nontrading
j=O
in the daily data variance estimator in the text. Adjustments were also

made for nontrading days (i.e., weekends and holidays) by dividing the

"daily" returns by the square root of the number of days between trades.
 ABl

Appendix B

The Posterior Distribution for Y.

J

In Section III, we have from (111.1) that

2 .
Xl (t) = Y. [a(t) 1 -J + e: (t) (B.l)
J

for j = 1,2,3 and t = 1, ... ,N where the prior density for Y.

J
is

uniform with f(Y.) = lib for 0 < Y. < b and zero, otherwise.
J - J-
Because the {e:(t)} are independently and identically distributed

standard normal, the joint density for XI(l), ••• ,XI(N), conditional on

(B.2)
1
- -2 QJ' 1 2 2 N/2
=e exp[~ -2 Q.(Y.-X.) 1/(2rr)
J J J .

and· n2.
~&J
and '\
I\j

are as defined in (111.4.1) and (111.4.2).

By Bayes Theorem, the posterior density for Y. given Xl (1) , ••• ,Xl (N)
J
can be written as, j = 1,2,3,

(B.3)
b
1 2
.= exp[- -2 Q.(Yj-X.) 2
1/ i 12
exp[- -2 Q'(Yi-Xj) 2 ldy .•
J . J . J J
o
 AB2

By the change in the variable of integration u = nj[Y j ~ Aj ], we have

that
p.

-[ { .- "21 u 2 du ] tn
j

(B.4)
= 2n {~(p.) - ~(n.)} In.
J J J

where Pj =n.(b
J
- A.)
J
and n.J =-A J.nJ.• By comb ining (B.3) and (B.4), we

have expression (111.3) for F[Y./ ] given in the text.

J
" 2
To determine Y. = E[Y . IY. , n. ; b] , we simply multiply F[Y.I ] by
J J J J J
Y. and integrate from Y.= 0 to Y. = b, and note that " = A.•
Y.
J J J J J
 Bl

Bibliography

1. Breeden, D., 1979. "An Intertemporal Asset Pricing Model With

Stochastic Consumption and Investment Opportunities," Research Paper
497, Graduate School of Business, Stanford University (July).

2. Black, F., 1976. "Studies of Stock Price Volatility Changes,"

Proceedings of the 1976 Meetings of Business and Economic Statistics
Section, American Statistical Association, pp. 177-181.

3. , M.C. Jensen, and M. Scholes, 1972. "The Capital Asset

Pricing Model: Some Empirical Tests,"'in Studies in the Theory of
Capital 11arkets, M.C. Jensen, ed., New York: Praeger.

4. , and M. Scholes, 1973. "The Pricing ·of Options and

Corporate Liabilities," Journal of Political Economy 81 (May-June):
pp. 637-59.

5. Blume, M. and I. Friend, 1975. "The Asset Structure of Individual

Portfolios and Some Implications for Utility Functions," Journal of
Finance 30 (May), pp. 585-603.

6. Cox, J., J. Ingersoll, and S.Ross, (forthcoming). "A Theory of

the Term Structure of Interest Rates," Econometrica.

7. Fama, E. and J. MacBeth, 1974. "Tests of the Multi-Period Two-

Parameter Model," Journal of Financial Economics 1, (May), pp. 43-66.

8. ~ .and G. Schwert, 1977. "Humail capital and capital Market

Equilibrium," Journal of Financial Economics 4, (January), pp. 95-
125.

9. Friend, I., and M. Blume, 1970. "Measurement of Portfolio

Performance Under Uncertainty," American Economic Review 60, pp.
561-575.

10. Ibbots<m~ R., and R. Sinquefield, 1976. "Stocks, Bonds, Bills, and
Inflation: Year-by-Year Historical Returns (1926-1974)," Journal
of Business 49 (January), pp. 11-47.

11. ,1977. Stocks, Bonds, Bills, and

Inflation: The Past (1926-1976) and The Future (1977-2000),
Charlottsville, Virginia: Financial Analysts Research Foundation.

12. ,i979. Stocks, Bonds, Bills, and

~I~n£~l-a-t~i-o-n-:~H=~~'s-t-o-r-~~'c-al~~R~e-t-u-r-n-s~(~1926-l978). Charlottesville,
Virginia: Financial Analysts Research Foundation.

13. Jensen, M.C., 1972. "Capital Markets: Theory and Evidence," Bell
Journal of Econ01llics and Management Science 3, pp. 357-398. --
 B2

14. Latane', H. and R. Rendleman, 1976. "Standard Deviations of Stock

Price Ratios Implied in Option Prices," Journal of Finance 31 (May),
pp. 369-381.

15. Lintner, J., 1965. "The Valuation of Risk Assets and The Selection
of Risky Investments in Stock Portfolios and Capital Budgets,"
Review of Economics and Statistics 47, pp. 13-37. (

16. Malkiel, B. and J. Cragg, 1979. "Expectations and The ValUation of

Shares" (mimeo).

17. Merton, R.C., 1971. "Optimum Consumption and Portfolio Rules in a

Continuous Time Model," Journal of Economic Theory 3 (December),
pp. 373-413.

18. , 1973. "An Intertemporal Capital Asset Pricing Model,"

Econometrica 41 (September), pp. 867-887.

19. , 1976. "The Impact on Option Pricing of Specification

Error in the Underlying Stock Returns," Journal of Finance 31 (May),
pp. 333-350.

20. , 1978. "On the "Mathematics and Economic Assumptions of

Continuous-Time Models," A.P. Sloan School of Management Working
Paper 11981-78, MIT (March).

21. , 1980. "On the Microeconomic Theory of Investment

Under Uncertainty," in Handbook of Hathematical Economics, K. Arrow
and M. Intriligator (eds.), North-Holland~

22. Mossin, J., 1966. "Equilibrium in a Capital Asset Market," Econo-

met>.ica 35, pp. 768-783.

23. Myers, S.C. and G. Pogue, 1979. "Measuring Expected Rates of Return,
Risk Premiums, and the Market Price of Risk" (mimeo).

24. Rosenberg, B., 1972. "The Behavior of Random Variables with Non-
stationary Variance and the Distribution of Security Prices,"
Graduate School of Business Administration Working Paper #11,
University of California, Berkeley (December).

25. Ross, S., 1976. "Arbitrage Theory of Capital Asset Pricing,"

. Journal of Economic Theory 8.

26. Rothschild, M. and J. Stiglitz, 1970. "Increasing Risk I: A

Definition," Journal of Economic Theory.2, pp. 225-243.

27. Schmalensee, R. and R. Trippi, 1978. "Common Stock Volatility

Expectations Implied by Option Premia," Journal of Finance 33,
(March},pp. 129-147.
 B3

28. Sharpe, W., 1964. "Capital Asset Prices: A Theory of Market

Equilibrium Under Conditions of Risk," Journal of Finance .19
(September), pp. 425-442.

29. Treynor, J. and F. Black, 1973. "How to Use Security Analysis to

Improve Portfolio Selection," Journal of Business 46 (January),
pp. 66-86.